This project provides a training framework for a 3D Faster R-CNN model specifically designed for lung nodule detection in CT images. It targets researchers and practitioners in medical imaging analysis, offering a novel approach to training deep convolutional networks for this task.

How It Works

The framework utilizes an Intel-extended Caffe version, optimized for Intel CPUs and supporting 3D convolutional layers. It combines elements from Faster R-CNN, U-Net, and ResNet architectures, incorporating a 150-layer deep convolutional network. This approach aims to achieve convergence and good performance on CT image data, a challenging domain for 3D CNNs.

Quick Start & Requirements

• Install: [sudo] pip install -r requirements.txt
• Prerequisites: Intel Extended Caffe (with MKLML engine and pycaffe installed), Python environment.
• Hardware: Recommended Intel CPUs (Xeon E series, Xeon Phi). Training on desktop PCs may be slow.
• Data: Requires preprocessing of CT images, including lung segmentation and label file creation. Links to preprocessing scripts and data preparation steps are provided.

Highlighted Details

• First training framework for 3D Faster R-CNN RPN with a 150-layer DCN on CT images.
• Achieved good performance on the Alibaba Tianchi Healthcare AI Competition data.
• Model architecture combines Faster R-CNN, U-Net, and ResNet blocks.
• Supports training and detection on unlabeled data.

Maintenance & Community

Developed by Shenzhen Yiyuan Intelligence Tech Co., LTD and Hong Kong Baptist University, GPU High Performance Computing Laboratory. No specific community links (Discord/Slack) or roadmap are mentioned.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility with commercial or closed-source projects is not specified.

Limitations & Caveats

The framework is CPU-based and explicitly welcomes PRs for GPU migration. The lung segmentation preprocessing method is noted as suboptimal, with potential for errors requiring manual thresholding. The project is described as a "training framework," implying it may not be a ready-to-deploy inference solution without further adaptation.